Add benchmarks for floating point math #618
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
tgross35
force-pushed
the
benchmarking
branch
8 times, most recently
from
4d58cdb
to
f1f9786
Compare
|
What is the reason for using global_asm? I would think that normal inline asm should be enough as long as you specify the constraints? Or is it because you want to accurately measure function call overhead as well? |
|
Normal inline assembly seemed to save rax even when not needed, https://rust.godbolt.org/z/8vfnWE84s, so I just used global to be closer to what LLVM generates. Am I just missing a constraint? |
|
Ah, |
tgross35
force-pushed
the
benchmarking
branch
6 times, most recently
from
eea497f
to
3cee637
Compare
|
Everything is passing now so I think this should be good. Here is the output of a full run on a ryzen 5900X: `cargo bench` outputGood to know that we are pretty close to the system implementations for most everything. Interesting that our f128 add/sub/mul routines appear to be 2x as fast, I wonder why |
|
Some assorted notes after looking at the graphs, just for reference:
|
This adds comparisons among the compiler-builtins function, system functions if available, and optionally handwritten assembly. These also help us identify inconsistencies between this crate and system functions, which may otherwise go unnoticed if intrinsics get lowered to inline operations rather than library calls.
|
On most CPUs, division instructions have variable latency depending on the input values. |
|
Enough to be slower than soft float on average? That seems surprising. Thanks for the review. |
|
The hardware division instruction blocks the pipeline, while the soft-float implementation is fully pipelined. This means that with a sufficiently wide CPU, the execution units could be running instructions from different iterations of the benchmark loop concurrently. On the other hand CPUs tend to only have 1 division unit, so no parallelism is possible. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This adds comparisons among the compiler-builtins function, system functions if available, and optionally handwritten assembly.
These also serve as some additional testing since we check our functions against assembly operations.
Run with
cargo bench --features benchmarking-reportsto get the HTML charts